Accurate forecasting of fine particulate matter (PM₂.₅) concentrations is important to protect public health. However, predicting PM₂.₅ accurately is difficult because pollution levels vary greatly over space and time due to weather, geography, and pollution sources. To improve forecasts, we developed an advanced artificial intelligence model combining Bidirectional Long Short-Term Memory networks and Double Kolmogorov-Arnold Networks. Our model uses hourly pollution data and weather conditions to capture complex temporal changes. It also considers spatial information such as nearby monitoring stations, population density, and distance to highways. We evaluated our approach using extensive pollution data from Texas. Results showed significant improvements compared to traditional forecasting methods. At a detailed 1 km spatial resolution, the model correctly identified high-pollution days more than 93% of the time. This improved approach can help environmental agencies provide timely pollution alerts and protect community health.